![Self-Supervised Surround-View Depth Estimation with Volumetric Feature Fusion](/assets/img/vision/fusion/volumetric_feature_fusion/0.png)
Self-Supervised Surround-View Depth Estimation with Volumetric Feature Fusion
2023, Mar 18
- 논문 : https://openreview.net/forum?id=0PfIQs-ttQQ
- 발표 자료 : https://nips.cc/virtual/2022/poster/54283
- 이번 글에서는
NIPS 2022
에 발표된Self-Supervised Surround-View Depth Estimation with Volumetric Feature Fusion
논문에 대한 내용 리뷰를 진행하겠습니다. - 논문에서 주목하고자 하는 부분은 멀티 카메라를 사용하였을 때 카메라 간 겹치는 영역이 발생하는데 그 영역에 대하여 어떻게 Fusion을 잘 할 지에 대한 방법과 그 효과를 보여줍니다. 이 방법을 논문에서는
Volumetric Feature Fusion
이라고 명하였습니다.
목차
-
Abstract
-
Introduction
-
Related Work
-
Surround-View Depth Estimation via Volumetric Feature Fusion
-
Experiments
-
Conclusions
-
Supplementary
-
42dot dataset
Abstract
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/2.png)
Introduction
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/3.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/4.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/5.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/6.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/7.gif)
Related Work
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/8.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/9.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/10.png)
Surround-View Depth Estimation via Volumetric Feature Fusion
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/11.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/12.png)
Architecture overview
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/13.png)
Surround-view volumetric feature fusion
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/14.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/15.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/16.png)
Depth Estimation
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/17.png)
Canonical motion estimation
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/18.png)
Self-supervised learning
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/19.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/20.png)
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/21.png)
Experiments
Conclusions
Supplementary
42dot dataset
- 아래 링크에서 42dot 데이터셋을 받아볼 수 있습니다.
- 링크 : https://www.42dot.ai/akit/dataset/mcmot
- 본 논문에서는
front
,front-left
,front-right
에 대한 시각화 사례가 있었는데, 42dot 데이터셋을 보면front
,front-left
,front-right
을 사용한 것을 볼 수 있습니다. 링크의 설명은 보면front
는 60도 화각의 카메라이며front-left
,front-right
는 120도 화각의 카메라임을 알 수 있습니다. - 아래는
volumetric feature fusion
을 한 영역으로 추정되는 영역을 표시하였습니다.
![Drawing](../assets/img/vision/fusion/volumetric_feature_fusion/1.png)